Enhancing Stereo Matching Domain Generalization With Adversarial Domain Alignment

This paper addresses a challenge where stereo-matching networks (AI systems that estimate depth by comparing two images) perform well on synthetic training data but struggle with real-world images due to domain gap (the difference between training and real-world data). The researchers propose ADASM, a method using adversarial domain alignment (exposing the model to worst-case scenarios during training to improve robustness) to make these networks generalize better to unseen real-world data without requiring fine-tuning.